There's been a recent call from @dynomight to declare whether and how you're using AI to write your essays. That seems reasonable, so here's my own policy for how I use AI in my blog:

AI can help, but I’m the author

• I use AI as an extensive aid for thinking, but retain primary authorship over my words. Unless I explicitly state otherwise, almost every sentence that's not explicitly a quote from an AI is written by me, not AI.
• The exception is that if I discuss a topic with an AI and it suggests a good point that I agree with, I may copy a sentence or two from it directly into the essay. I'm however sensitive to sentences that don't feel like they're in my own voice, and will generally rewrite them until they are.
  • Currently I only remember one essay from 1½ years ago where I included a direct quote that wasn't marked as a quote or edited at least a bit, because I found it too jarring to re-read later.
  • Any unchanged sentences in articles that are cross-posted to LW will be labeled, as per the LW policy on LLM use.
    • LW policy says that "text written by an LLM and then edited or revised by a human" should also be disclosed but that text written by a human "developed with LLM assistance" doesn't need to, unless one "borrows language" from the LLM. I never got an answer to my question of what "borrowing language" means and generally find it confusing to try to figure out when a sentence counts as something I've written vs. one that I've revised. So instead I use the "is it a direct quote" and "integrated contributions only" (below) tests, and expect these to meet the spirit of the policy. If they don't, sorry.
• I use AI (generally, Claude Opus or Claude Fable) for brainstorming, background research, checking for meanings of words and expressions, help with analogies and examples, for analysis and critique, and various smaller things.

Integrated contributions only

• When AI suggests a new idea, point, or framing, I try to only use it in an essay if it feels like one that genuinely integrates with my thinking.
• This is a bit fuzzy, but the test is something like: if a human had suggested this in a live conversation yesterday, would I still find the gist of what they said compelling even if I had forgotten their exact words? Even better, might I have built on it to add something that they didn’t say and might not even fully agree with?
• The opposite of this would be something that made me go “I don’t really understand that and I’d have to re-read it to even remember what exactly it said, but it sounds very cool and impressive”.
• Another test I have is: when I’m reading the draft before publishing it, does the writing flow in a natural way that feels “all genuinely grounded in me”? Or, are there sections where the idea came from AI, that break the flow and make the vibe fuzzy or vague?

Example AI contributions

• Pretty much all of my recent essays have undergone changes or gotten additions as a result of feedback or discussion from AI.
• The following list of examples is not exhaustive.
• In line with the “integrated contributions only” principle, all of these are ones that I could still recall directly, without having to re-read the essays in question:
  • An essay about how my experience of gender had changed originally included a line saying that "nothing had been lost" due to the changes I experienced. AI called that line out as having a slightly suspicious vibe. I thought about it and realized that I was indeed papering over sadness about something that felt lost, and changed that section to more honestly share what I felt.
  • In discussing an essay about how people's emotional needs corrupt the way they interpret ideas, AI suggested at least two additional points. First, that the way a given person corrupts any given idea might be predictable from their specific psychology. Second, that some frameworks might be ambiguous about what they're saying in a way that makes them spread further, as that ambiguity lets some people interpret the framework in a way that fits their needs and others in a way that fits theirs. Both of these ideas went into the essay.
  • An essay about why many therapies and spiritual frameworks claim to be The One True One originally included a list of examples of this, then my synthesis of what I think is really going on with people's emotional problems, and then an analysis of how people end up thinking they have the one true solution. AI convinced me that the second section was on much weaker ground than the others and not strictly needed for the final analysis, so I moved the synthesis to its own post and replaced it with a more general take that didn't require believing in all the details of my synthesis.
  • In the same essay, I reported an experience that I felt supported my position on something. AI pointed out a way in which the exactly same experience would be compatible with the opposite position. I added a footnote acknowledging the opposite interpretation and that just this experience wouldn't be decisive on its own.
  • In the same essay, I both quoted from a review article that AI found for me that supported my position, as well as adding links to studies that the AI pointed out which contradicted my position. I chose not to include some other articles that it found in support of my position, either because I could not access them to verify their contents, or because I did get to look at their contents and thought that they didn't quite say what the AI claimed they said.
  • In reviewing a draft of this very policy, AI pointed out that the opening emphasized sentence-level independence, but that the later examples showed significant AI influence on structure and content, and that this might mislead people who only read the opening and skimmed the rest. It suggested opening with a sentence like "AI shapes my thinking more than my prose" for clarity, which I then rephrased into "I use AI as an extensive aid for thinking, but retain primary authorship over my words" and added that.
  • AI also suggested moving this list of examples from the end to somewhere early on, which I did.

Specific use cases

Brainstorming

• If I have a rough idea that's not yet ready for an essay, I may discuss it with an AI to get preliminary reflections and pointers to relevant concepts.
• Sometimes this is pure rubber ducking (the chat format just gets me to write out my thoughts and I effectively ignore everything that the AI says in response), sometimes it gives me additional points to incorporate, sometimes it gets me to discard an idea completely when the AI points out the flaws in it.
• At other times I might approach it with a question rather than an idea, e.g. "here's something I'm struggling to make sense of", especially if it's something that I strongly believe or intend to build on in an essay, but also know of something strongly implying the opposite.

Background research

• If my writing e.g. touches upon something that people have studied, I ask AI to look into the relevant research. If I suspect there's probably an established term for a concept I'm developing, I ask the AI for pointers.
• For any empirical claims that are important for my post, I ask the AI for sources for its claims and I'll personally verify that they have been summarized correctly.

Checking for meanings of words and expressions

• If I think of using a certain kind of expression but I'm not totally sure I understand it correctly, I will ask an AI "what does X in this paragraph mean" and give it a paragraph where I use that expression to see if it interprets it the way I intend it to.

Help with analogies and examples

• Sometimes I have a rough idea for what I want to say, and will ask AI for ideas with that.
• Example of a recent prompt where I did this: "I'm writing an article about how some mental issues are ones where you pretty much need one specific intervention for them, while others can be tackled in multiple ways. I'm trying to make an analogy to conventional medicine that's something like "if you have a broken bone, putting it into a cast and letting it heal is pretty much the only correct approach. But if you have X, then either Y or Z can work." (Where some of the Ys and Zs can be something non-medical too, like "get more exercise".)"
• I only incorporate analogies that I feel like I understand well enough to stand behind. In this case, AI suggested type 2 diabetes, high blood pressure, and chronic back pain as a possible Xs. I decided to use high blood pressure specifically because I was more familiar with it than the other examples.

Analysis and critique

• I will give a finished draft to an AI and ask it to analyze it, then consider addressing any weaknesses or incorporating any new connections it points out.

Various smaller things

• I may ask AI to generate a graph whose contents I make sure are correct, sometimes ask it to suggest titles/subtitles for an essay, ask for help when I'm struggling to express a point, let tools like Grammarly suggest minor improvements to grammar or flow, etc.